Comprehensive Machine Learning Prediction of Extensive Enzymatic Reactions

Area of Science:

• Biochemistry and Biotechnology
• Computational Biology and Bioinformatics
• Metabolic Engineering

Background:

• A vast number of unannotated protein sequences contain undiscovered enzyme functions.
• Discovering novel enzymes is crucial for expanding metabolic engineering pathways for biosynthesis.
• Current machine learning models for enzymatic reaction prediction have limitations in predicting unknown reactions.

Purpose of the Study:

• To develop and evaluate expanded enzymatic reaction prediction models.
• To assess the capability of machine learning models, particularly deep neural networks, in predicting unknown enzymatic reactions.
• To improve the prediction accuracy and scope of enzymatic reactions for novel enzyme discovery.

Main Methods:

• Development of 16 expanded enzymatic reaction prediction models using diverse machine learning algorithms, including deep neural networks.
• Training models with combined substrate-enzyme-product information.
• Evaluation of prediction performance, including accuracy for known and unknown enzymatic reactions.

Main Results:

• The updated prediction models demonstrate improved performance compared to previous studies.
• The deep neural network model achieved the highest prediction accuracy, with Macro F1 scores up to 0.966.
• The deep neural network model robustly predicted unknown enzymatic reactions not present in the training data, outperforming existing models.

Conclusions:

• The developed deep neural network model significantly enhances the prediction of a wider range of enzymatic reactions.
• This advancement facilitates the discovery of novel enzymes with potential applications in biosynthesis.
• The study provides a powerful tool for advancing metabolic engineering and the production of valuable compounds.